drivers/md/bcache/super.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * bcache setup/teardown code, and some metadata io - read a superblock and
   4  * figure out what to do with it.
   5  *
   6  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   7  * Copyright 2012 Google, Inc.
   8  */
   9
  10 #include "bcache.h"
  11 #include "btree.h"
  12 #include "debug.h"
  13 #include "extents.h"
  14 #include "request.h"
  15 #include "writeback.h"
  16 #include "features.h"
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/debugfs.h>
  20 #include <linux/genhd.h>
  21 #include <linux/idr.h>
  22 #include <linux/kthread.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/module.h>
  25 #include <linux/random.h>
  26 #include <linux/reboot.h>
  27 #include <linux/sysfs.h>
  28
  29 unsigned int bch_cutoff_writeback;
  30 unsigned int bch_cutoff_writeback_sync;
  31
  32 static const char bcache_magic[] = {
  33         0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
  34         0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
  35 };
  36
  37 static const char invalid_uuid[] = {
  38         0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
  39         0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
  40 };
  41
  42 static struct kobject *bcache_kobj;
  43 struct mutex bch_register_lock;
  44 bool bcache_is_reboot;
  45 LIST_HEAD(bch_cache_sets);
  46 static LIST_HEAD(uncached_devices);
  47
  48 static int bcache_major;
  49 static DEFINE_IDA(bcache_device_idx);
  50 static wait_queue_head_t unregister_wait;
  51 struct workqueue_struct *bcache_wq;
  52 struct workqueue_struct *bch_flush_wq;
  53 struct workqueue_struct *bch_journal_wq;
  54
  55
  56 #define BTREE_MAX_PAGES         (256 * 1024 / PAGE_SIZE)
  57 /* limitation of partitions number on single bcache device */
  58 #define BCACHE_MINORS           128
  59 /* limitation of bcache devices number on single system */
  60 #define BCACHE_DEVICE_IDX_MAX   ((1U << MINORBITS)/BCACHE_MINORS)
  61
  62 /* Superblock */
  63
  64 static unsigned int get_bucket_size(struct cache_sb *sb, struct cache_sb_disk *s)
  65 {
  66         unsigned int bucket_size = le16_to_cpu(s->bucket_size);
  67
  68         if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
  69                 if (bch_has_feature_large_bucket(sb)) {
  70                         unsigned int max, order;
  71
  72                         max = sizeof(unsigned int) * BITS_PER_BYTE - 1;
  73                         order = le16_to_cpu(s->bucket_size);
  74                         /*
  75                          * bcache tool will make sure the overflow won't
  76                          * happen, an error message here is enough.
  77                          */
  78                         if (order > max)
  79                                 pr_err("Bucket size (1 << %u) overflows\n",
  80                                         order);
  81                         bucket_size = 1 << order;
  82                 } else if (bch_has_feature_obso_large_bucket(sb)) {
  83                         bucket_size +=
  84                                 le16_to_cpu(s->obso_bucket_size_hi) << 16;
  85                 }
  86         }
  87
  88         return bucket_size;
  89 }
  90
  91 static const char *read_super_common(struct cache_sb *sb,  struct block_device *bdev,
  92                                      struct cache_sb_disk *s)
  93 {
  94         const char *err;
  95         unsigned int i;
  96
  97         sb->first_bucket= le16_to_cpu(s->first_bucket);
  98         sb->nbuckets    = le64_to_cpu(s->nbuckets);
  99         sb->bucket_size = get_bucket_size(sb, s);
 100
 101         sb->nr_in_set   = le16_to_cpu(s->nr_in_set);
 102         sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
 103
 104         err = "Too many journal buckets";
 105         if (sb->keys > SB_JOURNAL_BUCKETS)
 106                 goto err;
 107
 108         err = "Too many buckets";
 109         if (sb->nbuckets > LONG_MAX)
 110                 goto err;
 111
 112         err = "Not enough buckets";
 113         if (sb->nbuckets < 1 << 7)
 114                 goto err;
 115
 116         err = "Bad block size (not power of 2)";
 117         if (!is_power_of_2(sb->block_size))
 118                 goto err;
 119
 120         err = "Bad block size (larger than page size)";
 121         if (sb->block_size > PAGE_SECTORS)
 122                 goto err;
 123
 124         err = "Bad bucket size (not power of 2)";
 125         if (!is_power_of_2(sb->bucket_size))
 126                 goto err;
 127
 128         err = "Bad bucket size (smaller than page size)";
 129         if (sb->bucket_size < PAGE_SECTORS)
 130                 goto err;
 131
 132         err = "Invalid superblock: device too small";
 133         if (get_capacity(bdev->bd_disk) <
 134             sb->bucket_size * sb->nbuckets)
 135                 goto err;
 136
 137         err = "Bad UUID";
 138         if (bch_is_zero(sb->set_uuid, 16))
 139                 goto err;
 140
 141         err = "Bad cache device number in set";
 142         if (!sb->nr_in_set ||
 143             sb->nr_in_set <= sb->nr_this_dev ||
 144             sb->nr_in_set > MAX_CACHES_PER_SET)
 145                 goto err;
 146
 147         err = "Journal buckets not sequential";
 148         for (i = 0; i < sb->keys; i++)
 149                 if (sb->d[i] != sb->first_bucket + i)
 150                         goto err;
 151
 152         err = "Too many journal buckets";
 153         if (sb->first_bucket + sb->keys > sb->nbuckets)
 154                 goto err;
 155
 156         err = "Invalid superblock: first bucket comes before end of super";
 157         if (sb->first_bucket * sb->bucket_size < 16)
 158                 goto err;
 159
 160         err = NULL;
 161 err:
 162         return err;
 163 }
 164
 165
 166 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
 167                               struct cache_sb_disk **res)
 168 {
 169         const char *err;
 170         struct cache_sb_disk *s;
 171         struct page *page;
 172         unsigned int i;
 173
 174         page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
 175                                    SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
 176         if (IS_ERR(page))
 177                 return "IO error";
 178         s = page_address(page) + offset_in_page(SB_OFFSET);
 179
 180         sb->offset              = le64_to_cpu(s->offset);
 181         sb->version             = le64_to_cpu(s->version);
 182
 183         memcpy(sb->magic,       s->magic, 16);
 184         memcpy(sb->uuid,        s->uuid, 16);
 185         memcpy(sb->set_uuid,    s->set_uuid, 16);
 186         memcpy(sb->label,       s->label, SB_LABEL_SIZE);
 187
 188         sb->flags               = le64_to_cpu(s->flags);
 189         sb->seq                 = le64_to_cpu(s->seq);
 190         sb->last_mount          = le32_to_cpu(s->last_mount);
 191         sb->keys                = le16_to_cpu(s->keys);
 192
 193         for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
 194                 sb->d[i] = le64_to_cpu(s->d[i]);
 195
 196         pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
 197                  sb->version, sb->flags, sb->seq, sb->keys);
 198
 199         err = "Not a bcache superblock (bad offset)";
 200         if (sb->offset != SB_SECTOR)
 201                 goto err;
 202
 203         err = "Not a bcache superblock (bad magic)";
 204         if (memcmp(sb->magic, bcache_magic, 16))
 205                 goto err;
 206
 207         err = "Bad checksum";
 208         if (s->csum != csum_set(s))
 209                 goto err;
 210
 211         err = "Bad UUID";
 212         if (bch_is_zero(sb->uuid, 16))
 213                 goto err;
 214
 215         sb->block_size  = le16_to_cpu(s->block_size);
 216
 217         err = "Superblock block size smaller than device block size";
 218         if (sb->block_size << 9 < bdev_logical_block_size(bdev))
 219                 goto err;
 220
 221         switch (sb->version) {
 222         case BCACHE_SB_VERSION_BDEV:
 223                 sb->data_offset = BDEV_DATA_START_DEFAULT;
 224                 break;
 225         case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
 226         case BCACHE_SB_VERSION_BDEV_WITH_FEATURES:
 227                 sb->data_offset = le64_to_cpu(s->data_offset);
 228
 229                 err = "Bad data offset";
 230                 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
 231                         goto err;
 232
 233                 break;
 234         case BCACHE_SB_VERSION_CDEV:
 235         case BCACHE_SB_VERSION_CDEV_WITH_UUID:
 236                 err = read_super_common(sb, bdev, s);
 237                 if (err)
 238                         goto err;
 239                 break;
 240         case BCACHE_SB_VERSION_CDEV_WITH_FEATURES:
 241                 /*
 242                  * Feature bits are needed in read_super_common(),
 243                  * convert them firstly.
 244                  */
 245                 sb->feature_compat = le64_to_cpu(s->feature_compat);
 246                 sb->feature_incompat = le64_to_cpu(s->feature_incompat);
 247                 sb->feature_ro_compat = le64_to_cpu(s->feature_ro_compat);
 248
 249                 /* Check incompatible features */
 250                 err = "Unsupported compatible feature found";
 251                 if (bch_has_unknown_compat_features(sb))
 252                         goto err;
 253
 254                 err = "Unsupported read-only compatible feature found";
 255                 if (bch_has_unknown_ro_compat_features(sb))
 256                         goto err;
 257
 258                 err = "Unsupported incompatible feature found";
 259                 if (bch_has_unknown_incompat_features(sb))
 260                         goto err;
 261
 262                 err = read_super_common(sb, bdev, s);
 263                 if (err)
 264                         goto err;
 265                 break;
 266         default:
 267                 err = "Unsupported superblock version";
 268                 goto err;
 269         }
 270
 271         sb->last_mount = (u32)ktime_get_real_seconds();
 272         *res = s;
 273         return NULL;
 274 err:
 275         put_page(page);
 276         return err;
 277 }
 278
 279 static void write_bdev_super_endio(struct bio *bio)
 280 {
 281         struct cached_dev *dc = bio->bi_private;
 282
 283         if (bio->bi_status)
 284                 bch_count_backing_io_errors(dc, bio);
 285
 286         closure_put(&dc->sb_write);
 287 }
 288
 289 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
 290                 struct bio *bio)
 291 {
 292         unsigned int i;
 293
 294         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
 295         bio->bi_iter.bi_sector  = SB_SECTOR;
 296         __bio_add_page(bio, virt_to_page(out), SB_SIZE,
 297                         offset_in_page(out));
 298
 299         out->offset             = cpu_to_le64(sb->offset);
 300
 301         memcpy(out->uuid,       sb->uuid, 16);
 302         memcpy(out->set_uuid,   sb->set_uuid, 16);
 303         memcpy(out->label,      sb->label, SB_LABEL_SIZE);
 304
 305         out->flags              = cpu_to_le64(sb->flags);
 306         out->seq                = cpu_to_le64(sb->seq);
 307
 308         out->last_mount         = cpu_to_le32(sb->last_mount);
 309         out->first_bucket       = cpu_to_le16(sb->first_bucket);
 310         out->keys               = cpu_to_le16(sb->keys);
 311
 312         for (i = 0; i < sb->keys; i++)
 313                 out->d[i] = cpu_to_le64(sb->d[i]);
 314
 315         if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
 316                 out->feature_compat    = cpu_to_le64(sb->feature_compat);
 317                 out->feature_incompat  = cpu_to_le64(sb->feature_incompat);
 318                 out->feature_ro_compat = cpu_to_le64(sb->feature_ro_compat);
 319         }
 320
 321         out->version            = cpu_to_le64(sb->version);
 322         out->csum = csum_set(out);
 323
 324         pr_debug("ver %llu, flags %llu, seq %llu\n",
 325                  sb->version, sb->flags, sb->seq);
 326
 327         submit_bio(bio);
 328 }
 329
 330 static void bch_write_bdev_super_unlock(struct closure *cl)
 331 {
 332         struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
 333
 334         up(&dc->sb_write_mutex);
 335 }
 336
 337 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
 338 {
 339         struct closure *cl = &dc->sb_write;
 340         struct bio *bio = &dc->sb_bio;
 341
 342         down(&dc->sb_write_mutex);
 343         closure_init(cl, parent);
 344
 345         bio_init(bio, dc->sb_bv, 1);
 346         bio_set_dev(bio, dc->bdev);
 347         bio->bi_end_io  = write_bdev_super_endio;
 348         bio->bi_private = dc;
 349
 350         closure_get(cl);
 351         /* I/O request sent to backing device */
 352         __write_super(&dc->sb, dc->sb_disk, bio);
 353
 354         closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
 355 }
 356
 357 static void write_super_endio(struct bio *bio)
 358 {
 359         struct cache *ca = bio->bi_private;
 360
 361         /* is_read = 0 */
 362         bch_count_io_errors(ca, bio->bi_status, 0,
 363                             "writing superblock");
 364         closure_put(&ca->set->sb_write);
 365 }
 366
 367 static void bcache_write_super_unlock(struct closure *cl)
 368 {
 369         struct cache_set *c = container_of(cl, struct cache_set, sb_write);
 370
 371         up(&c->sb_write_mutex);
 372 }
 373
 374 void bcache_write_super(struct cache_set *c)
 375 {
 376         struct closure *cl = &c->sb_write;
 377         struct cache *ca = c->cache;
 378         struct bio *bio = &ca->sb_bio;
 379         unsigned int version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
 380
 381         down(&c->sb_write_mutex);
 382         closure_init(cl, &c->cl);
 383
 384         ca->sb.seq++;
 385
 386         if (ca->sb.version < version)
 387                 ca->sb.version = version;
 388
 389         bio_init(bio, ca->sb_bv, 1);
 390         bio_set_dev(bio, ca->bdev);
 391         bio->bi_end_io  = write_super_endio;
 392         bio->bi_private = ca;
 393
 394         closure_get(cl);
 395         __write_super(&ca->sb, ca->sb_disk, bio);
 396
 397         closure_return_with_destructor(cl, bcache_write_super_unlock);
 398 }
 399
 400 /* UUID io */
 401
 402 static void uuid_endio(struct bio *bio)
 403 {
 404         struct closure *cl = bio->bi_private;
 405         struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
 406
 407         cache_set_err_on(bio->bi_status, c, "accessing uuids");
 408         bch_bbio_free(bio, c);
 409         closure_put(cl);
 410 }
 411
 412 static void uuid_io_unlock(struct closure *cl)
 413 {
 414         struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
 415
 416         up(&c->uuid_write_mutex);
 417 }
 418
 419 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
 420                     struct bkey *k, struct closure *parent)
 421 {
 422         struct closure *cl = &c->uuid_write;
 423         struct uuid_entry *u;
 424         unsigned int i;
 425         char buf[80];
 426
 427         BUG_ON(!parent);
 428         down(&c->uuid_write_mutex);
 429         closure_init(cl, parent);
 430
 431         for (i = 0; i < KEY_PTRS(k); i++) {
 432                 struct bio *bio = bch_bbio_alloc(c);
 433
 434                 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
 435                 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
 436
 437                 bio->bi_end_io  = uuid_endio;
 438                 bio->bi_private = cl;
 439                 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
 440                 bch_bio_map(bio, c->uuids);
 441
 442                 bch_submit_bbio(bio, c, k, i);
 443
 444                 if (op != REQ_OP_WRITE)
 445                         break;
 446         }
 447
 448         bch_extent_to_text(buf, sizeof(buf), k);
 449         pr_debug("%s UUIDs at %s\n", op == REQ_OP_WRITE ? "wrote" : "read", buf);
 450
 451         for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
 452                 if (!bch_is_zero(u->uuid, 16))
 453                         pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u\n",
 454                                  u - c->uuids, u->uuid, u->label,
 455                                  u->first_reg, u->last_reg, u->invalidated);
 456
 457         closure_return_with_destructor(cl, uuid_io_unlock);
 458 }
 459
 460 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
 461 {
 462         struct bkey *k = &j->uuid_bucket;
 463
 464         if (__bch_btree_ptr_invalid(c, k))
 465                 return "bad uuid pointer";
 466
 467         bkey_copy(&c->uuid_bucket, k);
 468         uuid_io(c, REQ_OP_READ, 0, k, cl);
 469
 470         if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
 471                 struct uuid_entry_v0    *u0 = (void *) c->uuids;
 472                 struct uuid_entry       *u1 = (void *) c->uuids;
 473                 int i;
 474
 475                 closure_sync(cl);
 476
 477                 /*
 478                  * Since the new uuid entry is bigger than the old, we have to
 479                  * convert starting at the highest memory address and work down
 480                  * in order to do it in place
 481                  */
 482
 483                 for (i = c->nr_uuids - 1;
 484                      i >= 0;
 485                      --i) {
 486                         memcpy(u1[i].uuid,      u0[i].uuid, 16);
 487                         memcpy(u1[i].label,     u0[i].label, 32);
 488
 489                         u1[i].first_reg         = u0[i].first_reg;
 490                         u1[i].last_reg          = u0[i].last_reg;
 491                         u1[i].invalidated       = u0[i].invalidated;
 492
 493                         u1[i].flags     = 0;
 494                         u1[i].sectors   = 0;
 495                 }
 496         }
 497
 498         return NULL;
 499 }
 500
 501 static int __uuid_write(struct cache_set *c)
 502 {
 503         BKEY_PADDED(key) k;
 504         struct closure cl;
 505         struct cache *ca = c->cache;
 506         unsigned int size;
 507
 508         closure_init_stack(&cl);
 509         lockdep_assert_held(&bch_register_lock);
 510
 511         if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, true))
 512                 return 1;
 513
 514         size =  meta_bucket_pages(&ca->sb) * PAGE_SECTORS;
 515         SET_KEY_SIZE(&k.key, size);
 516         uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
 517         closure_sync(&cl);
 518
 519         /* Only one bucket used for uuid write */
 520         atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
 521
 522         bkey_copy(&c->uuid_bucket, &k.key);
 523         bkey_put(c, &k.key);
 524         return 0;
 525 }
 526
 527 int bch_uuid_write(struct cache_set *c)
 528 {
 529         int ret = __uuid_write(c);
 530
 531         if (!ret)
 532                 bch_journal_meta(c, NULL);
 533
 534         return ret;
 535 }
 536
 537 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
 538 {
 539         struct uuid_entry *u;
 540
 541         for (u = c->uuids;
 542              u < c->uuids + c->nr_uuids; u++)
 543                 if (!memcmp(u->uuid, uuid, 16))
 544                         return u;
 545
 546         return NULL;
 547 }
 548
 549 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
 550 {
 551         static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
 552
 553         return uuid_find(c, zero_uuid);
 554 }
 555
 556 /*
 557  * Bucket priorities/gens:
 558  *
 559  * For each bucket, we store on disk its
 560  *   8 bit gen
 561  *  16 bit priority
 562  *
 563  * See alloc.c for an explanation of the gen. The priority is used to implement
 564  * lru (and in the future other) cache replacement policies; for most purposes
 565  * it's just an opaque integer.
 566  *
 567  * The gens and the priorities don't have a whole lot to do with each other, and
 568  * it's actually the gens that must be written out at specific times - it's no
 569  * big deal if the priorities don't get written, if we lose them we just reuse
 570  * buckets in suboptimal order.
 571  *
 572  * On disk they're stored in a packed array, and in as many buckets are required
 573  * to fit them all. The buckets we use to store them form a list; the journal
 574  * header points to the first bucket, the first bucket points to the second
 575  * bucket, et cetera.
 576  *
 577  * This code is used by the allocation code; periodically (whenever it runs out
 578  * of buckets to allocate from) the allocation code will invalidate some
 579  * buckets, but it can't use those buckets until their new gens are safely on
 580  * disk.
 581  */
 582
 583 static void prio_endio(struct bio *bio)
 584 {
 585         struct cache *ca = bio->bi_private;
 586
 587         cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
 588         bch_bbio_free(bio, ca->set);
 589         closure_put(&ca->prio);
 590 }
 591
 592 static void prio_io(struct cache *ca, uint64_t bucket, int op,
 593                     unsigned long op_flags)
 594 {
 595         struct closure *cl = &ca->prio;
 596         struct bio *bio = bch_bbio_alloc(ca->set);
 597
 598         closure_init_stack(cl);
 599
 600         bio->bi_iter.bi_sector  = bucket * ca->sb.bucket_size;
 601         bio_set_dev(bio, ca->bdev);
 602         bio->bi_iter.bi_size    = meta_bucket_bytes(&ca->sb);
 603
 604         bio->bi_end_io  = prio_endio;
 605         bio->bi_private = ca;
 606         bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
 607         bch_bio_map(bio, ca->disk_buckets);
 608
 609         closure_bio_submit(ca->set, bio, &ca->prio);
 610         closure_sync(cl);
 611 }
 612
 613 int bch_prio_write(struct cache *ca, bool wait)
 614 {
 615         int i;
 616         struct bucket *b;
 617         struct closure cl;
 618
 619         pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
 620                  fifo_used(&ca->free[RESERVE_PRIO]),
 621                  fifo_used(&ca->free[RESERVE_NONE]),
 622                  fifo_used(&ca->free_inc));
 623
 624         /*
 625          * Pre-check if there are enough free buckets. In the non-blocking
 626          * scenario it's better to fail early rather than starting to allocate
 627          * buckets and do a cleanup later in case of failure.
 628          */
 629         if (!wait) {
 630                 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
 631                                fifo_used(&ca->free[RESERVE_NONE]);
 632                 if (prio_buckets(ca) > avail)
 633                         return -ENOMEM;
 634         }
 635
 636         closure_init_stack(&cl);
 637
 638         lockdep_assert_held(&ca->set->bucket_lock);
 639
 640         ca->disk_buckets->seq++;
 641
 642         atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
 643                         &ca->meta_sectors_written);
 644
 645         for (i = prio_buckets(ca) - 1; i >= 0; --i) {
 646                 long bucket;
 647                 struct prio_set *p = ca->disk_buckets;
 648                 struct bucket_disk *d = p->data;
 649                 struct bucket_disk *end = d + prios_per_bucket(ca);
 650
 651                 for (b = ca->buckets + i * prios_per_bucket(ca);
 652                      b < ca->buckets + ca->sb.nbuckets && d < end;
 653                      b++, d++) {
 654                         d->prio = cpu_to_le16(b->prio);
 655                         d->gen = b->gen;
 656                 }
 657
 658                 p->next_bucket  = ca->prio_buckets[i + 1];
 659                 p->magic        = pset_magic(&ca->sb);
 660                 p->csum         = bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8);
 661
 662                 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
 663                 BUG_ON(bucket == -1);
 664
 665                 mutex_unlock(&ca->set->bucket_lock);
 666                 prio_io(ca, bucket, REQ_OP_WRITE, 0);
 667                 mutex_lock(&ca->set->bucket_lock);
 668
 669                 ca->prio_buckets[i] = bucket;
 670                 atomic_dec_bug(&ca->buckets[bucket].pin);
 671         }
 672
 673         mutex_unlock(&ca->set->bucket_lock);
 674
 675         bch_journal_meta(ca->set, &cl);
 676         closure_sync(&cl);
 677
 678         mutex_lock(&ca->set->bucket_lock);
 679
 680         /*
 681          * Don't want the old priorities to get garbage collected until after we
 682          * finish writing the new ones, and they're journalled
 683          */
 684         for (i = 0; i < prio_buckets(ca); i++) {
 685                 if (ca->prio_last_buckets[i])
 686                         __bch_bucket_free(ca,
 687                                 &ca->buckets[ca->prio_last_buckets[i]]);
 688
 689                 ca->prio_last_buckets[i] = ca->prio_buckets[i];
 690         }
 691         return 0;
 692 }
 693
 694 static int prio_read(struct cache *ca, uint64_t bucket)
 695 {
 696         struct prio_set *p = ca->disk_buckets;
 697         struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
 698         struct bucket *b;
 699         unsigned int bucket_nr = 0;
 700         int ret = -EIO;
 701
 702         for (b = ca->buckets;
 703              b < ca->buckets + ca->sb.nbuckets;
 704              b++, d++) {
 705                 if (d == end) {
 706                         ca->prio_buckets[bucket_nr] = bucket;
 707                         ca->prio_last_buckets[bucket_nr] = bucket;
 708                         bucket_nr++;
 709
 710                         prio_io(ca, bucket, REQ_OP_READ, 0);
 711
 712                         if (p->csum !=
 713                             bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8)) {
 714                                 pr_warn("bad csum reading priorities\n");
 715                                 goto out;
 716                         }
 717
 718                         if (p->magic != pset_magic(&ca->sb)) {
 719                                 pr_warn("bad magic reading priorities\n");
 720                                 goto out;
 721                         }
 722
 723                         bucket = p->next_bucket;
 724                         d = p->data;
 725                 }
 726
 727                 b->prio = le16_to_cpu(d->prio);
 728                 b->gen = b->last_gc = d->gen;
 729         }
 730
 731         ret = 0;
 732 out:
 733         return ret;
 734 }
 735
 736 /* Bcache device */
 737
 738 static int open_dev(struct block_device *b, fmode_t mode)
 739 {
 740         struct bcache_device *d = b->bd_disk->private_data;
 741
 742         if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
 743                 return -ENXIO;
 744
 745         closure_get(&d->cl);
 746         return 0;
 747 }
 748
 749 static void release_dev(struct gendisk *b, fmode_t mode)
 750 {
 751         struct bcache_device *d = b->private_data;
 752
 753         closure_put(&d->cl);
 754 }
 755
 756 static int ioctl_dev(struct block_device *b, fmode_t mode,
 757                      unsigned int cmd, unsigned long arg)
 758 {
 759         struct bcache_device *d = b->bd_disk->private_data;
 760
 761         return d->ioctl(d, mode, cmd, arg);
 762 }
 763
 764 static const struct block_device_operations bcache_cached_ops = {
 765         .submit_bio     = cached_dev_submit_bio,
 766         .open           = open_dev,
 767         .release        = release_dev,
 768         .ioctl          = ioctl_dev,
 769         .owner          = THIS_MODULE,
 770 };
 771
 772 static const struct block_device_operations bcache_flash_ops = {
 773         .submit_bio     = flash_dev_submit_bio,
 774         .open           = open_dev,
 775         .release        = release_dev,
 776         .ioctl          = ioctl_dev,
 777         .owner          = THIS_MODULE,
 778 };
 779
 780 void bcache_device_stop(struct bcache_device *d)
 781 {
 782         if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
 783                 /*
 784                  * closure_fn set to
 785                  * - cached device: cached_dev_flush()
 786                  * - flash dev: flash_dev_flush()
 787                  */
 788                 closure_queue(&d->cl);
 789 }
 790
 791 static void bcache_device_unlink(struct bcache_device *d)
 792 {
 793         lockdep_assert_held(&bch_register_lock);
 794
 795         if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
 796                 struct cache *ca = d->c->cache;
 797
 798                 sysfs_remove_link(&d->c->kobj, d->name);
 799                 sysfs_remove_link(&d->kobj, "cache");
 800
 801                 bd_unlink_disk_holder(ca->bdev, d->disk);
 802         }
 803 }
 804
 805 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
 806                                const char *name)
 807 {
 808         struct cache *ca = c->cache;
 809         int ret;
 810
 811         bd_link_disk_holder(ca->bdev, d->disk);
 812
 813         snprintf(d->name, BCACHEDEVNAME_SIZE,
 814                  "%s%u", name, d->id);
 815
 816         ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
 817         if (ret < 0)
 818                 pr_err("Couldn't create device -> cache set symlink\n");
 819
 820         ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
 821         if (ret < 0)
 822                 pr_err("Couldn't create cache set -> device symlink\n");
 823
 824         clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
 825 }
 826
 827 static void bcache_device_detach(struct bcache_device *d)
 828 {
 829         lockdep_assert_held(&bch_register_lock);
 830
 831         atomic_dec(&d->c->attached_dev_nr);
 832
 833         if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
 834                 struct uuid_entry *u = d->c->uuids + d->id;
 835
 836                 SET_UUID_FLASH_ONLY(u, 0);
 837                 memcpy(u->uuid, invalid_uuid, 16);
 838                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
 839                 bch_uuid_write(d->c);
 840         }
 841
 842         bcache_device_unlink(d);
 843
 844         d->c->devices[d->id] = NULL;
 845         closure_put(&d->c->caching);
 846         d->c = NULL;
 847 }
 848
 849 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
 850                                  unsigned int id)
 851 {
 852         d->id = id;
 853         d->c = c;
 854         c->devices[id] = d;
 855
 856         if (id >= c->devices_max_used)
 857                 c->devices_max_used = id + 1;
 858
 859         closure_get(&c->caching);
 860 }
 861
 862 static inline int first_minor_to_idx(int first_minor)
 863 {
 864         return (first_minor/BCACHE_MINORS);
 865 }
 866
 867 static inline int idx_to_first_minor(int idx)
 868 {
 869         return (idx * BCACHE_MINORS);
 870 }
 871
 872 static void bcache_device_free(struct bcache_device *d)
 873 {
 874         struct gendisk *disk = d->disk;
 875
 876         lockdep_assert_held(&bch_register_lock);
 877
 878         if (disk)
 879                 pr_info("%s stopped\n", disk->disk_name);
 880         else
 881                 pr_err("bcache device (NULL gendisk) stopped\n");
 882
 883         if (d->c)
 884                 bcache_device_detach(d);
 885
 886         if (disk) {
 887                 bool disk_added = (disk->flags & GENHD_FL_UP) != 0;
 888
 889                 if (disk_added)
 890                         del_gendisk(disk);
 891
 892                 if (disk->queue)
 893                         blk_cleanup_queue(disk->queue);
 894
 895                 ida_simple_remove(&bcache_device_idx,
 896                                   first_minor_to_idx(disk->first_minor));
 897                 if (disk_added)
 898                         put_disk(disk);
 899         }
 900
 901         bioset_exit(&d->bio_split);
 902         kvfree(d->full_dirty_stripes);
 903         kvfree(d->stripe_sectors_dirty);
 904
 905         closure_debug_destroy(&d->cl);
 906 }
 907
 908 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
 909                 sector_t sectors, struct block_device *cached_bdev,
 910                 const struct block_device_operations *ops)
 911 {
 912         struct request_queue *q;
 913         const size_t max_stripes = min_t(size_t, INT_MAX,
 914                                          SIZE_MAX / sizeof(atomic_t));
 915         uint64_t n;
 916         int idx;
 917
 918         if (!d->stripe_size)
 919                 d->stripe_size = 1 << 31;
 920
 921         n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
 922         if (!n || n > max_stripes) {
 923                 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
 924                         n);
 925                 return -ENOMEM;
 926         }
 927         d->nr_stripes = n;
 928
 929         n = d->nr_stripes * sizeof(atomic_t);
 930         d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
 931         if (!d->stripe_sectors_dirty)
 932                 return -ENOMEM;
 933
 934         n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
 935         d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
 936         if (!d->full_dirty_stripes)
 937                 return -ENOMEM;
 938
 939         idx = ida_simple_get(&bcache_device_idx, 0,
 940                                 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
 941         if (idx < 0)
 942                 return idx;
 943
 944         if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
 945                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
 946                 goto err;
 947
 948         d->disk = alloc_disk(BCACHE_MINORS);
 949         if (!d->disk)
 950                 goto err;
 951
 952         set_capacity(d->disk, sectors);
 953         snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
 954
 955         d->disk->major          = bcache_major;
 956         d->disk->first_minor    = idx_to_first_minor(idx);
 957         d->disk->fops           = ops;
 958         d->disk->private_data   = d;
 959
 960         q = blk_alloc_queue(NUMA_NO_NODE);
 961         if (!q)
 962                 return -ENOMEM;
 963
 964         d->disk->queue                  = q;
 965         q->limits.max_hw_sectors        = UINT_MAX;
 966         q->limits.max_sectors           = UINT_MAX;
 967         q->limits.max_segment_size      = UINT_MAX;
 968         q->limits.max_segments          = BIO_MAX_VECS;
 969         blk_queue_max_discard_sectors(q, UINT_MAX);
 970         q->limits.discard_granularity   = 512;
 971         q->limits.io_min                = block_size;
 972         q->limits.logical_block_size    = block_size;
 973         q->limits.physical_block_size   = block_size;
 974
 975         if (q->limits.logical_block_size > PAGE_SIZE && cached_bdev) {
 976                 /*
 977                  * This should only happen with BCACHE_SB_VERSION_BDEV.
 978                  * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
 979                  */
 980                 pr_info("%s: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
 981                         d->disk->disk_name, q->limits.logical_block_size,
 982                         PAGE_SIZE, bdev_logical_block_size(cached_bdev));
 983
 984                 /* This also adjusts physical block size/min io size if needed */
 985                 blk_queue_logical_block_size(q, bdev_logical_block_size(cached_bdev));
 986         }
 987
 988         blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
 989         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
 990         blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
 991
 992         blk_queue_write_cache(q, true, true);
 993
 994         return 0;
 995
 996 err:
 997         ida_simple_remove(&bcache_device_idx, idx);
 998         return -ENOMEM;
 999
1000 }
1001
1002 /* Cached device */
1003
1004 static void calc_cached_dev_sectors(struct cache_set *c)
1005 {
1006         uint64_t sectors = 0;
1007         struct cached_dev *dc;
1008
1009         list_for_each_entry(dc, &c->cached_devs, list)
1010                 sectors += bdev_sectors(dc->bdev);
1011
1012         c->cached_dev_sectors = sectors;
1013 }
1014
1015 #define BACKING_DEV_OFFLINE_TIMEOUT 5
1016 static int cached_dev_status_update(void *arg)
1017 {
1018         struct cached_dev *dc = arg;
1019         struct request_queue *q;
1020
1021         /*
1022          * If this delayed worker is stopping outside, directly quit here.
1023          * dc->io_disable might be set via sysfs interface, so check it
1024          * here too.
1025          */
1026         while (!kthread_should_stop() && !dc->io_disable) {
1027                 q = bdev_get_queue(dc->bdev);
1028                 if (blk_queue_dying(q))
1029                         dc->offline_seconds++;
1030                 else
1031                         dc->offline_seconds = 0;
1032
1033                 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1034                         pr_err("%s: device offline for %d seconds\n",
1035                                dc->backing_dev_name,
1036                                BACKING_DEV_OFFLINE_TIMEOUT);
1037                         pr_err("%s: disable I/O request due to backing device offline\n",
1038                                dc->disk.name);
1039                         dc->io_disable = true;
1040                         /* let others know earlier that io_disable is true */
1041                         smp_mb();
1042                         bcache_device_stop(&dc->disk);
1043                         break;
1044                 }
1045                 schedule_timeout_interruptible(HZ);
1046         }
1047
1048         wait_for_kthread_stop();
1049         return 0;
1050 }
1051
1052
1053 int bch_cached_dev_run(struct cached_dev *dc)
1054 {
1055         int ret = 0;
1056         struct bcache_device *d = &dc->disk;
1057         char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1058         char *env[] = {
1059                 "DRIVER=bcache",
1060                 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1061                 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1062                 NULL,
1063         };
1064
1065         if (dc->io_disable) {
1066                 pr_err("I/O disabled on cached dev %s\n",
1067                        dc->backing_dev_name);
1068                 ret = -EIO;
1069                 goto out;
1070         }
1071
1072         if (atomic_xchg(&dc->running, 1)) {
1073                 pr_info("cached dev %s is running already\n",
1074                        dc->backing_dev_name);
1075                 ret = -EBUSY;
1076                 goto out;
1077         }
1078
1079         if (!d->c &&
1080             BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1081                 struct closure cl;
1082
1083                 closure_init_stack(&cl);
1084
1085                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1086                 bch_write_bdev_super(dc, &cl);
1087                 closure_sync(&cl);
1088         }
1089
1090         add_disk(d->disk);
1091         bd_link_disk_holder(dc->bdev, dc->disk.disk);
1092         /*
1093          * won't show up in the uevent file, use udevadm monitor -e instead
1094          * only class / kset properties are persistent
1095          */
1096         kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1097
1098         if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1099             sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1100                               &d->kobj, "bcache")) {
1101                 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1102                 ret = -ENOMEM;
1103                 goto out;
1104         }
1105
1106         dc->status_update_thread = kthread_run(cached_dev_status_update,
1107                                                dc, "bcache_status_update");
1108         if (IS_ERR(dc->status_update_thread)) {
1109                 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1110         }
1111
1112 out:
1113         kfree(env[1]);
1114         kfree(env[2]);
1115         kfree(buf);
1116         return ret;
1117 }
1118
1119 /*
1120  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1121  * work dc->writeback_rate_update is running. Wait until the routine
1122  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1123  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1124  * seconds, give up waiting here and continue to cancel it too.
1125  */
1126 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1127 {
1128         int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1129
1130         do {
1131                 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1132                               &dc->disk.flags))
1133                         break;
1134                 time_out--;
1135                 schedule_timeout_interruptible(1);
1136         } while (time_out > 0);
1137
1138         if (time_out == 0)
1139                 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1140
1141         cancel_delayed_work_sync(&dc->writeback_rate_update);
1142 }
1143
1144 static void cached_dev_detach_finish(struct work_struct *w)
1145 {
1146         struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1147
1148         BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1149         BUG_ON(refcount_read(&dc->count));
1150
1151
1152         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1153                 cancel_writeback_rate_update_dwork(dc);
1154
1155         if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1156                 kthread_stop(dc->writeback_thread);
1157                 dc->writeback_thread = NULL;
1158         }
1159
1160         mutex_lock(&bch_register_lock);
1161
1162         calc_cached_dev_sectors(dc->disk.c);
1163         bcache_device_detach(&dc->disk);
1164         list_move(&dc->list, &uncached_devices);
1165
1166         clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1167         clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1168
1169         mutex_unlock(&bch_register_lock);
1170
1171         pr_info("Caching disabled for %s\n", dc->backing_dev_name);
1172
1173         /* Drop ref we took in cached_dev_detach() */
1174         closure_put(&dc->disk.cl);
1175 }
1176
1177 void bch_cached_dev_detach(struct cached_dev *dc)
1178 {
1179         lockdep_assert_held(&bch_register_lock);
1180
1181         if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1182                 return;
1183
1184         if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1185                 return;
1186
1187         /*
1188          * Block the device from being closed and freed until we're finished
1189          * detaching
1190          */
1191         closure_get(&dc->disk.cl);
1192
1193         bch_writeback_queue(dc);
1194
1195         cached_dev_put(dc);
1196 }
1197
1198 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1199                           uint8_t *set_uuid)
1200 {
1201         uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1202         struct uuid_entry *u;
1203         struct cached_dev *exist_dc, *t;
1204         int ret = 0;
1205
1206         if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1207             (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1208                 return -ENOENT;
1209
1210         if (dc->disk.c) {
1211                 pr_err("Can't attach %s: already attached\n",
1212                        dc->backing_dev_name);
1213                 return -EINVAL;
1214         }
1215
1216         if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1217                 pr_err("Can't attach %s: shutting down\n",
1218                        dc->backing_dev_name);
1219                 return -EINVAL;
1220         }
1221
1222         if (dc->sb.block_size < c->cache->sb.block_size) {
1223                 /* Will die */
1224                 pr_err("Couldn't attach %s: block size less than set's block size\n",
1225                        dc->backing_dev_name);
1226                 return -EINVAL;
1227         }
1228
1229         /* Check whether already attached */
1230         list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1231                 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1232                         pr_err("Tried to attach %s but duplicate UUID already attached\n",
1233                                 dc->backing_dev_name);
1234
1235                         return -EINVAL;
1236                 }
1237         }
1238
1239         u = uuid_find(c, dc->sb.uuid);
1240
1241         if (u &&
1242             (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1243              BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1244                 memcpy(u->uuid, invalid_uuid, 16);
1245                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1246                 u = NULL;
1247         }
1248
1249         if (!u) {
1250                 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1251                         pr_err("Couldn't find uuid for %s in set\n",
1252                                dc->backing_dev_name);
1253                         return -ENOENT;
1254                 }
1255
1256                 u = uuid_find_empty(c);
1257                 if (!u) {
1258                         pr_err("Not caching %s, no room for UUID\n",
1259                                dc->backing_dev_name);
1260                         return -EINVAL;
1261                 }
1262         }
1263
1264         /*
1265          * Deadlocks since we're called via sysfs...
1266          * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1267          */
1268
1269         if (bch_is_zero(u->uuid, 16)) {
1270                 struct closure cl;
1271
1272                 closure_init_stack(&cl);
1273
1274                 memcpy(u->uuid, dc->sb.uuid, 16);
1275                 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1276                 u->first_reg = u->last_reg = rtime;
1277                 bch_uuid_write(c);
1278
1279                 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1280                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1281
1282                 bch_write_bdev_super(dc, &cl);
1283                 closure_sync(&cl);
1284         } else {
1285                 u->last_reg = rtime;
1286                 bch_uuid_write(c);
1287         }
1288
1289         bcache_device_attach(&dc->disk, c, u - c->uuids);
1290         list_move(&dc->list, &c->cached_devs);
1291         calc_cached_dev_sectors(c);
1292
1293         /*
1294          * dc->c must be set before dc->count != 0 - paired with the mb in
1295          * cached_dev_get()
1296          */
1297         smp_wmb();
1298         refcount_set(&dc->count, 1);
1299
1300         /* Block writeback thread, but spawn it */
1301         down_write(&dc->writeback_lock);
1302         if (bch_cached_dev_writeback_start(dc)) {
1303                 up_write(&dc->writeback_lock);
1304                 pr_err("Couldn't start writeback facilities for %s\n",
1305                        dc->disk.disk->disk_name);
1306                 return -ENOMEM;
1307         }
1308
1309         if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1310                 atomic_set(&dc->has_dirty, 1);
1311                 bch_writeback_queue(dc);
1312         }
1313
1314         bch_sectors_dirty_init(&dc->disk);
1315
1316         ret = bch_cached_dev_run(dc);
1317         if (ret && (ret != -EBUSY)) {
1318                 up_write(&dc->writeback_lock);
1319                 /*
1320                  * bch_register_lock is held, bcache_device_stop() is not
1321                  * able to be directly called. The kthread and kworker
1322                  * created previously in bch_cached_dev_writeback_start()
1323                  * have to be stopped manually here.
1324                  */
1325                 kthread_stop(dc->writeback_thread);
1326                 cancel_writeback_rate_update_dwork(dc);
1327                 pr_err("Couldn't run cached device %s\n",
1328                        dc->backing_dev_name);
1329                 return ret;
1330         }
1331
1332         bcache_device_link(&dc->disk, c, "bdev");
1333         atomic_inc(&c->attached_dev_nr);
1334
1335         if (bch_has_feature_obso_large_bucket(&(c->cache->sb))) {
1336                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1337                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1338                 set_disk_ro(dc->disk.disk, 1);
1339         }
1340
1341         /* Allow the writeback thread to proceed */
1342         up_write(&dc->writeback_lock);
1343
1344         pr_info("Caching %s as %s on set %pU\n",
1345                 dc->backing_dev_name,
1346                 dc->disk.disk->disk_name,
1347                 dc->disk.c->set_uuid);
1348         return 0;
1349 }
1350
1351 /* when dc->disk.kobj released */
1352 void bch_cached_dev_release(struct kobject *kobj)
1353 {
1354         struct cached_dev *dc = container_of(kobj, struct cached_dev,
1355                                              disk.kobj);
1356         kfree(dc);
1357         module_put(THIS_MODULE);
1358 }
1359
1360 static void cached_dev_free(struct closure *cl)
1361 {
1362         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1363
1364         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1365                 cancel_writeback_rate_update_dwork(dc);
1366
1367         if (!IS_ERR_OR_NULL(dc->writeback_thread))
1368                 kthread_stop(dc->writeback_thread);
1369         if (!IS_ERR_OR_NULL(dc->status_update_thread))
1370                 kthread_stop(dc->status_update_thread);
1371
1372         mutex_lock(&bch_register_lock);
1373
1374         if (atomic_read(&dc->running))
1375                 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1376         bcache_device_free(&dc->disk);
1377         list_del(&dc->list);
1378
1379         mutex_unlock(&bch_register_lock);
1380
1381         if (dc->sb_disk)
1382                 put_page(virt_to_page(dc->sb_disk));
1383
1384         if (!IS_ERR_OR_NULL(dc->bdev))
1385                 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1386
1387         wake_up(&unregister_wait);
1388
1389         kobject_put(&dc->disk.kobj);
1390 }
1391
1392 static void cached_dev_flush(struct closure *cl)
1393 {
1394         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1395         struct bcache_device *d = &dc->disk;
1396
1397         mutex_lock(&bch_register_lock);
1398         bcache_device_unlink(d);
1399         mutex_unlock(&bch_register_lock);
1400
1401         bch_cache_accounting_destroy(&dc->accounting);
1402         kobject_del(&d->kobj);
1403
1404         continue_at(cl, cached_dev_free, system_wq);
1405 }
1406
1407 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1408 {
1409         int ret;
1410         struct io *io;
1411         struct request_queue *q = bdev_get_queue(dc->bdev);
1412
1413         __module_get(THIS_MODULE);
1414         INIT_LIST_HEAD(&dc->list);
1415         closure_init(&dc->disk.cl, NULL);
1416         set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1417         kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1418         INIT_WORK(&dc->detach, cached_dev_detach_finish);
1419         sema_init(&dc->sb_write_mutex, 1);
1420         INIT_LIST_HEAD(&dc->io_lru);
1421         spin_lock_init(&dc->io_lock);
1422         bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1423
1424         dc->sequential_cutoff           = 4 << 20;
1425
1426         for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1427                 list_add(&io->lru, &dc->io_lru);
1428                 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1429         }
1430
1431         dc->disk.stripe_size = q->limits.io_opt >> 9;
1432
1433         if (dc->disk.stripe_size)
1434                 dc->partial_stripes_expensive =
1435                         q->limits.raid_partial_stripes_expensive;
1436
1437         ret = bcache_device_init(&dc->disk, block_size,
1438                          bdev_nr_sectors(dc->bdev) - dc->sb.data_offset,
1439                          dc->bdev, &bcache_cached_ops);
1440         if (ret)
1441                 return ret;
1442
1443         blk_queue_io_opt(dc->disk.disk->queue,
1444                 max(queue_io_opt(dc->disk.disk->queue), queue_io_opt(q)));
1445
1446         atomic_set(&dc->io_errors, 0);
1447         dc->io_disable = false;
1448         dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1449         /* default to auto */
1450         dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1451
1452         bch_cached_dev_request_init(dc);
1453         bch_cached_dev_writeback_init(dc);
1454         return 0;
1455 }
1456
1457 /* Cached device - bcache superblock */
1458
1459 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1460                                  struct block_device *bdev,
1461                                  struct cached_dev *dc)
1462 {
1463         const char *err = "cannot allocate memory";
1464         struct cache_set *c;
1465         int ret = -ENOMEM;
1466
1467         bdevname(bdev, dc->backing_dev_name);
1468         memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1469         dc->bdev = bdev;
1470         dc->bdev->bd_holder = dc;
1471         dc->sb_disk = sb_disk;
1472
1473         if (cached_dev_init(dc, sb->block_size << 9))
1474                 goto err;
1475
1476         err = "error creating kobject";
1477         if (kobject_add(&dc->disk.kobj, bdev_kobj(bdev), "bcache"))
1478                 goto err;
1479         if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1480                 goto err;
1481
1482         pr_info("registered backing device %s\n", dc->backing_dev_name);
1483
1484         list_add(&dc->list, &uncached_devices);
1485         /* attach to a matched cache set if it exists */
1486         list_for_each_entry(c, &bch_cache_sets, list)
1487                 bch_cached_dev_attach(dc, c, NULL);
1488
1489         if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1490             BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1491                 err = "failed to run cached device";
1492                 ret = bch_cached_dev_run(dc);
1493                 if (ret)
1494                         goto err;
1495         }
1496
1497         return 0;
1498 err:
1499         pr_notice("error %s: %s\n", dc->backing_dev_name, err);
1500         bcache_device_stop(&dc->disk);
1501         return ret;
1502 }
1503
1504 /* Flash only volumes */
1505
1506 /* When d->kobj released */
1507 void bch_flash_dev_release(struct kobject *kobj)
1508 {
1509         struct bcache_device *d = container_of(kobj, struct bcache_device,
1510                                                kobj);
1511         kfree(d);
1512 }
1513
1514 static void flash_dev_free(struct closure *cl)
1515 {
1516         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1517
1518         mutex_lock(&bch_register_lock);
1519         atomic_long_sub(bcache_dev_sectors_dirty(d),
1520                         &d->c->flash_dev_dirty_sectors);
1521         bcache_device_free(d);
1522         mutex_unlock(&bch_register_lock);
1523         kobject_put(&d->kobj);
1524 }
1525
1526 static void flash_dev_flush(struct closure *cl)
1527 {
1528         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1529
1530         mutex_lock(&bch_register_lock);
1531         bcache_device_unlink(d);
1532         mutex_unlock(&bch_register_lock);
1533         kobject_del(&d->kobj);
1534         continue_at(cl, flash_dev_free, system_wq);
1535 }
1536
1537 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1538 {
1539         struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1540                                           GFP_KERNEL);
1541         if (!d)
1542                 return -ENOMEM;
1543
1544         closure_init(&d->cl, NULL);
1545         set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1546
1547         kobject_init(&d->kobj, &bch_flash_dev_ktype);
1548
1549         if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1550                         NULL, &bcache_flash_ops))
1551                 goto err;
1552
1553         bcache_device_attach(d, c, u - c->uuids);
1554         bch_sectors_dirty_init(d);
1555         bch_flash_dev_request_init(d);
1556         add_disk(d->disk);
1557
1558         if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1559                 goto err;
1560
1561         bcache_device_link(d, c, "volume");
1562
1563         if (bch_has_feature_obso_large_bucket(&c->cache->sb)) {
1564                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1565                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1566                 set_disk_ro(d->disk, 1);
1567         }
1568
1569         return 0;
1570 err:
1571         kobject_put(&d->kobj);
1572         return -ENOMEM;
1573 }
1574
1575 static int flash_devs_run(struct cache_set *c)
1576 {
1577         int ret = 0;
1578         struct uuid_entry *u;
1579
1580         for (u = c->uuids;
1581              u < c->uuids + c->nr_uuids && !ret;
1582              u++)
1583                 if (UUID_FLASH_ONLY(u))
1584                         ret = flash_dev_run(c, u);
1585
1586         return ret;
1587 }
1588
1589 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1590 {
1591         struct uuid_entry *u;
1592
1593         if (test_bit(CACHE_SET_STOPPING, &c->flags))
1594                 return -EINTR;
1595
1596         if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1597                 return -EPERM;
1598
1599         u = uuid_find_empty(c);
1600         if (!u) {
1601                 pr_err("Can't create volume, no room for UUID\n");
1602                 return -EINVAL;
1603         }
1604
1605         get_random_bytes(u->uuid, 16);
1606         memset(u->label, 0, 32);
1607         u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1608
1609         SET_UUID_FLASH_ONLY(u, 1);
1610         u->sectors = size >> 9;
1611
1612         bch_uuid_write(c);
1613
1614         return flash_dev_run(c, u);
1615 }
1616
1617 bool bch_cached_dev_error(struct cached_dev *dc)
1618 {
1619         if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1620                 return false;
1621
1622         dc->io_disable = true;
1623         /* make others know io_disable is true earlier */
1624         smp_mb();
1625
1626         pr_err("stop %s: too many IO errors on backing device %s\n",
1627                dc->disk.disk->disk_name, dc->backing_dev_name);
1628
1629         bcache_device_stop(&dc->disk);
1630         return true;
1631 }
1632
1633 /* Cache set */
1634
1635 __printf(2, 3)
1636 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1637 {
1638         struct va_format vaf;
1639         va_list args;
1640
1641         if (c->on_error != ON_ERROR_PANIC &&
1642             test_bit(CACHE_SET_STOPPING, &c->flags))
1643                 return false;
1644
1645         if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1646                 pr_info("CACHE_SET_IO_DISABLE already set\n");
1647
1648         /*
1649          * XXX: we can be called from atomic context
1650          * acquire_console_sem();
1651          */
1652
1653         va_start(args, fmt);
1654
1655         vaf.fmt = fmt;
1656         vaf.va = &args;
1657
1658         pr_err("error on %pU: %pV, disabling caching\n",
1659                c->set_uuid, &vaf);
1660
1661         va_end(args);
1662
1663         if (c->on_error == ON_ERROR_PANIC)
1664                 panic("panic forced after error\n");
1665
1666         bch_cache_set_unregister(c);
1667         return true;
1668 }
1669
1670 /* When c->kobj released */
1671 void bch_cache_set_release(struct kobject *kobj)
1672 {
1673         struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1674
1675         kfree(c);
1676         module_put(THIS_MODULE);
1677 }
1678
1679 static void cache_set_free(struct closure *cl)
1680 {
1681         struct cache_set *c = container_of(cl, struct cache_set, cl);
1682         struct cache *ca;
1683
1684         debugfs_remove(c->debug);
1685
1686         bch_open_buckets_free(c);
1687         bch_btree_cache_free(c);
1688         bch_journal_free(c);
1689
1690         mutex_lock(&bch_register_lock);
1691         bch_bset_sort_state_free(&c->sort);
1692         free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->cache->sb)));
1693
1694         ca = c->cache;
1695         if (ca) {
1696                 ca->set = NULL;
1697                 c->cache = NULL;
1698                 kobject_put(&ca->kobj);
1699         }
1700
1701
1702         if (c->moving_gc_wq)
1703                 destroy_workqueue(c->moving_gc_wq);
1704         bioset_exit(&c->bio_split);
1705         mempool_exit(&c->fill_iter);
1706         mempool_exit(&c->bio_meta);
1707         mempool_exit(&c->search);
1708         kfree(c->devices);
1709
1710         list_del(&c->list);
1711         mutex_unlock(&bch_register_lock);
1712
1713         pr_info("Cache set %pU unregistered\n", c->set_uuid);
1714         wake_up(&unregister_wait);
1715
1716         closure_debug_destroy(&c->cl);
1717         kobject_put(&c->kobj);
1718 }
1719
1720 static void cache_set_flush(struct closure *cl)
1721 {
1722         struct cache_set *c = container_of(cl, struct cache_set, caching);
1723         struct cache *ca = c->cache;
1724         struct btree *b;
1725
1726         bch_cache_accounting_destroy(&c->accounting);
1727
1728         kobject_put(&c->internal);
1729         kobject_del(&c->kobj);
1730
1731         if (!IS_ERR_OR_NULL(c->gc_thread))
1732                 kthread_stop(c->gc_thread);
1733
1734         if (!IS_ERR_OR_NULL(c->root))
1735                 list_add(&c->root->list, &c->btree_cache);
1736
1737         /*
1738          * Avoid flushing cached nodes if cache set is retiring
1739          * due to too many I/O errors detected.
1740          */
1741         if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1742                 list_for_each_entry(b, &c->btree_cache, list) {
1743                         mutex_lock(&b->write_lock);
1744                         if (btree_node_dirty(b))
1745                                 __bch_btree_node_write(b, NULL);
1746                         mutex_unlock(&b->write_lock);
1747                 }
1748
1749         if (ca->alloc_thread)
1750                 kthread_stop(ca->alloc_thread);
1751
1752         if (c->journal.cur) {
1753                 cancel_delayed_work_sync(&c->journal.work);
1754                 /* flush last journal entry if needed */
1755                 c->journal.work.work.func(&c->journal.work.work);
1756         }
1757
1758         closure_return(cl);
1759 }
1760
1761 /*
1762  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1763  * cache set is unregistering due to too many I/O errors. In this condition,
1764  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1765  * value and whether the broken cache has dirty data:
1766  *
1767  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1768  *  BCH_CACHED_STOP_AUTO               0               NO
1769  *  BCH_CACHED_STOP_AUTO               1               YES
1770  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1771  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1772  *
1773  * The expected behavior is, if stop_when_cache_set_failed is configured to
1774  * "auto" via sysfs interface, the bcache device will not be stopped if the
1775  * backing device is clean on the broken cache device.
1776  */
1777 static void conditional_stop_bcache_device(struct cache_set *c,
1778                                            struct bcache_device *d,
1779                                            struct cached_dev *dc)
1780 {
1781         if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1782                 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1783                         d->disk->disk_name, c->set_uuid);
1784                 bcache_device_stop(d);
1785         } else if (atomic_read(&dc->has_dirty)) {
1786                 /*
1787                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1788                  * and dc->has_dirty == 1
1789                  */
1790                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1791                         d->disk->disk_name);
1792                 /*
1793                  * There might be a small time gap that cache set is
1794                  * released but bcache device is not. Inside this time
1795                  * gap, regular I/O requests will directly go into
1796                  * backing device as no cache set attached to. This
1797                  * behavior may also introduce potential inconsistence
1798                  * data in writeback mode while cache is dirty.
1799                  * Therefore before calling bcache_device_stop() due
1800                  * to a broken cache device, dc->io_disable should be
1801                  * explicitly set to true.
1802                  */
1803                 dc->io_disable = true;
1804                 /* make others know io_disable is true earlier */
1805                 smp_mb();
1806                 bcache_device_stop(d);
1807         } else {
1808                 /*
1809                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1810                  * and dc->has_dirty == 0
1811                  */
1812                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1813                         d->disk->disk_name);
1814         }
1815 }
1816
1817 static void __cache_set_unregister(struct closure *cl)
1818 {
1819         struct cache_set *c = container_of(cl, struct cache_set, caching);
1820         struct cached_dev *dc;
1821         struct bcache_device *d;
1822         size_t i;
1823
1824         mutex_lock(&bch_register_lock);
1825
1826         for (i = 0; i < c->devices_max_used; i++) {
1827                 d = c->devices[i];
1828                 if (!d)
1829                         continue;
1830
1831                 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1832                     test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1833                         dc = container_of(d, struct cached_dev, disk);
1834                         bch_cached_dev_detach(dc);
1835                         if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1836                                 conditional_stop_bcache_device(c, d, dc);
1837                 } else {
1838                         bcache_device_stop(d);
1839                 }
1840         }
1841
1842         mutex_unlock(&bch_register_lock);
1843
1844         continue_at(cl, cache_set_flush, system_wq);
1845 }
1846
1847 void bch_cache_set_stop(struct cache_set *c)
1848 {
1849         if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1850                 /* closure_fn set to __cache_set_unregister() */
1851                 closure_queue(&c->caching);
1852 }
1853
1854 void bch_cache_set_unregister(struct cache_set *c)
1855 {
1856         set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1857         bch_cache_set_stop(c);
1858 }
1859
1860 #define alloc_meta_bucket_pages(gfp, sb)                \
1861         ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1862
1863 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1864 {
1865         int iter_size;
1866         struct cache *ca = container_of(sb, struct cache, sb);
1867         struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1868
1869         if (!c)
1870                 return NULL;
1871
1872         __module_get(THIS_MODULE);
1873         closure_init(&c->cl, NULL);
1874         set_closure_fn(&c->cl, cache_set_free, system_wq);
1875
1876         closure_init(&c->caching, &c->cl);
1877         set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1878
1879         /* Maybe create continue_at_noreturn() and use it here? */
1880         closure_set_stopped(&c->cl);
1881         closure_put(&c->cl);
1882
1883         kobject_init(&c->kobj, &bch_cache_set_ktype);
1884         kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1885
1886         bch_cache_accounting_init(&c->accounting, &c->cl);
1887
1888         memcpy(c->set_uuid, sb->set_uuid, 16);
1889
1890         c->cache                = ca;
1891         c->cache->set           = c;
1892         c->bucket_bits          = ilog2(sb->bucket_size);
1893         c->block_bits           = ilog2(sb->block_size);
1894         c->nr_uuids             = meta_bucket_bytes(sb) / sizeof(struct uuid_entry);
1895         c->devices_max_used     = 0;
1896         atomic_set(&c->attached_dev_nr, 0);
1897         c->btree_pages          = meta_bucket_pages(sb);
1898         if (c->btree_pages > BTREE_MAX_PAGES)
1899                 c->btree_pages = max_t(int, c->btree_pages / 4,
1900                                        BTREE_MAX_PAGES);
1901
1902         sema_init(&c->sb_write_mutex, 1);
1903         mutex_init(&c->bucket_lock);
1904         init_waitqueue_head(&c->btree_cache_wait);
1905         spin_lock_init(&c->btree_cannibalize_lock);
1906         init_waitqueue_head(&c->bucket_wait);
1907         init_waitqueue_head(&c->gc_wait);
1908         sema_init(&c->uuid_write_mutex, 1);
1909
1910         spin_lock_init(&c->btree_gc_time.lock);
1911         spin_lock_init(&c->btree_split_time.lock);
1912         spin_lock_init(&c->btree_read_time.lock);
1913
1914         bch_moving_init_cache_set(c);
1915
1916         INIT_LIST_HEAD(&c->list);
1917         INIT_LIST_HEAD(&c->cached_devs);
1918         INIT_LIST_HEAD(&c->btree_cache);
1919         INIT_LIST_HEAD(&c->btree_cache_freeable);
1920         INIT_LIST_HEAD(&c->btree_cache_freed);
1921         INIT_LIST_HEAD(&c->data_buckets);
1922
1923         iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size + 1) *
1924                 sizeof(struct btree_iter_set);
1925
1926         c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1927         if (!c->devices)
1928                 goto err;
1929
1930         if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1931                 goto err;
1932
1933         if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1934                         sizeof(struct bbio) +
1935                         sizeof(struct bio_vec) * meta_bucket_pages(sb)))
1936                 goto err;
1937
1938         if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1939                 goto err;
1940
1941         if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1942                         BIOSET_NEED_RESCUER))
1943                 goto err;
1944
1945         c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, sb);
1946         if (!c->uuids)
1947                 goto err;
1948
1949         c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1950         if (!c->moving_gc_wq)
1951                 goto err;
1952
1953         if (bch_journal_alloc(c))
1954                 goto err;
1955
1956         if (bch_btree_cache_alloc(c))
1957                 goto err;
1958
1959         if (bch_open_buckets_alloc(c))
1960                 goto err;
1961
1962         if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1963                 goto err;
1964
1965         c->congested_read_threshold_us  = 2000;
1966         c->congested_write_threshold_us = 20000;
1967         c->error_limit  = DEFAULT_IO_ERROR_LIMIT;
1968         c->idle_max_writeback_rate_enabled = 1;
1969         WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1970
1971         return c;
1972 err:
1973         bch_cache_set_unregister(c);
1974         return NULL;
1975 }
1976
1977 static int run_cache_set(struct cache_set *c)
1978 {
1979         const char *err = "cannot allocate memory";
1980         struct cached_dev *dc, *t;
1981         struct cache *ca = c->cache;
1982         struct closure cl;
1983         LIST_HEAD(journal);
1984         struct journal_replay *l;
1985
1986         closure_init_stack(&cl);
1987
1988         c->nbuckets = ca->sb.nbuckets;
1989         set_gc_sectors(c);
1990
1991         if (CACHE_SYNC(&c->cache->sb)) {
1992                 struct bkey *k;
1993                 struct jset *j;
1994
1995                 err = "cannot allocate memory for journal";
1996                 if (bch_journal_read(c, &journal))
1997                         goto err;
1998
1999                 pr_debug("btree_journal_read() done\n");
2000
2001                 err = "no journal entries found";
2002                 if (list_empty(&journal))
2003                         goto err;
2004
2005                 j = &list_entry(journal.prev, struct journal_replay, list)->j;
2006
2007                 err = "IO error reading priorities";
2008                 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
2009                         goto err;
2010
2011                 /*
2012                  * If prio_read() fails it'll call cache_set_error and we'll
2013                  * tear everything down right away, but if we perhaps checked
2014                  * sooner we could avoid journal replay.
2015                  */
2016
2017                 k = &j->btree_root;
2018
2019                 err = "bad btree root";
2020                 if (__bch_btree_ptr_invalid(c, k))
2021                         goto err;
2022
2023                 err = "error reading btree root";
2024                 c->root = bch_btree_node_get(c, NULL, k,
2025                                              j->btree_level,
2026                                              true, NULL);
2027                 if (IS_ERR_OR_NULL(c->root))
2028                         goto err;
2029
2030                 list_del_init(&c->root->list);
2031                 rw_unlock(true, c->root);
2032
2033                 err = uuid_read(c, j, &cl);
2034                 if (err)
2035                         goto err;
2036
2037                 err = "error in recovery";
2038                 if (bch_btree_check(c))
2039                         goto err;
2040
2041                 bch_journal_mark(c, &journal);
2042                 bch_initial_gc_finish(c);
2043                 pr_debug("btree_check() done\n");
2044
2045                 /*
2046                  * bcache_journal_next() can't happen sooner, or
2047                  * btree_gc_finish() will give spurious errors about last_gc >
2048                  * gc_gen - this is a hack but oh well.
2049                  */
2050                 bch_journal_next(&c->journal);
2051
2052                 err = "error starting allocator thread";
2053                 if (bch_cache_allocator_start(ca))
2054                         goto err;
2055
2056                 /*
2057                  * First place it's safe to allocate: btree_check() and
2058                  * btree_gc_finish() have to run before we have buckets to
2059                  * allocate, and bch_bucket_alloc_set() might cause a journal
2060                  * entry to be written so bcache_journal_next() has to be called
2061                  * first.
2062                  *
2063                  * If the uuids were in the old format we have to rewrite them
2064                  * before the next journal entry is written:
2065                  */
2066                 if (j->version < BCACHE_JSET_VERSION_UUID)
2067                         __uuid_write(c);
2068
2069                 err = "bcache: replay journal failed";
2070                 if (bch_journal_replay(c, &journal))
2071                         goto err;
2072         } else {
2073                 unsigned int j;
2074
2075                 pr_notice("invalidating existing data\n");
2076                 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2077                                         2, SB_JOURNAL_BUCKETS);
2078
2079                 for (j = 0; j < ca->sb.keys; j++)
2080                         ca->sb.d[j] = ca->sb.first_bucket + j;
2081
2082                 bch_initial_gc_finish(c);
2083
2084                 err = "error starting allocator thread";
2085                 if (bch_cache_allocator_start(ca))
2086                         goto err;
2087
2088                 mutex_lock(&c->bucket_lock);
2089                 bch_prio_write(ca, true);
2090                 mutex_unlock(&c->bucket_lock);
2091
2092                 err = "cannot allocate new UUID bucket";
2093                 if (__uuid_write(c))
2094                         goto err;
2095
2096                 err = "cannot allocate new btree root";
2097                 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2098                 if (IS_ERR_OR_NULL(c->root))
2099                         goto err;
2100
2101                 mutex_lock(&c->root->write_lock);
2102                 bkey_copy_key(&c->root->key, &MAX_KEY);
2103                 bch_btree_node_write(c->root, &cl);
2104                 mutex_unlock(&c->root->write_lock);
2105
2106                 bch_btree_set_root(c->root);
2107                 rw_unlock(true, c->root);
2108
2109                 /*
2110                  * We don't want to write the first journal entry until
2111                  * everything is set up - fortunately journal entries won't be
2112                  * written until the SET_CACHE_SYNC() here:
2113                  */
2114                 SET_CACHE_SYNC(&c->cache->sb, true);
2115
2116                 bch_journal_next(&c->journal);
2117                 bch_journal_meta(c, &cl);
2118         }
2119
2120         err = "error starting gc thread";
2121         if (bch_gc_thread_start(c))
2122                 goto err;
2123
2124         closure_sync(&cl);
2125         c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
2126         bcache_write_super(c);
2127
2128         if (bch_has_feature_obso_large_bucket(&c->cache->sb))
2129                 pr_err("Detect obsoleted large bucket layout, all attached bcache device will be read-only\n");
2130
2131         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2132                 bch_cached_dev_attach(dc, c, NULL);
2133
2134         flash_devs_run(c);
2135
2136         set_bit(CACHE_SET_RUNNING, &c->flags);
2137         return 0;
2138 err:
2139         while (!list_empty(&journal)) {
2140                 l = list_first_entry(&journal, struct journal_replay, list);
2141                 list_del(&l->list);
2142                 kfree(l);
2143         }
2144
2145         closure_sync(&cl);
2146
2147         bch_cache_set_error(c, "%s", err);
2148
2149         return -EIO;
2150 }
2151
2152 static const char *register_cache_set(struct cache *ca)
2153 {
2154         char buf[12];
2155         const char *err = "cannot allocate memory";
2156         struct cache_set *c;
2157
2158         list_for_each_entry(c, &bch_cache_sets, list)
2159                 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2160                         if (c->cache)
2161                                 return "duplicate cache set member";
2162
2163                         goto found;
2164                 }
2165
2166         c = bch_cache_set_alloc(&ca->sb);
2167         if (!c)
2168                 return err;
2169
2170         err = "error creating kobject";
2171         if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->set_uuid) ||
2172             kobject_add(&c->internal, &c->kobj, "internal"))
2173                 goto err;
2174
2175         if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2176                 goto err;
2177
2178         bch_debug_init_cache_set(c);
2179
2180         list_add(&c->list, &bch_cache_sets);
2181 found:
2182         sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2183         if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2184             sysfs_create_link(&c->kobj, &ca->kobj, buf))
2185                 goto err;
2186
2187         kobject_get(&ca->kobj);
2188         ca->set = c;
2189         ca->set->cache = ca;
2190
2191         err = "failed to run cache set";
2192         if (run_cache_set(c) < 0)
2193                 goto err;
2194
2195         return NULL;
2196 err:
2197         bch_cache_set_unregister(c);
2198         return err;
2199 }
2200
2201 /* Cache device */
2202
2203 /* When ca->kobj released */
2204 void bch_cache_release(struct kobject *kobj)
2205 {
2206         struct cache *ca = container_of(kobj, struct cache, kobj);
2207         unsigned int i;
2208
2209         if (ca->set) {
2210                 BUG_ON(ca->set->cache != ca);
2211                 ca->set->cache = NULL;
2212         }
2213
2214         free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2215         kfree(ca->prio_buckets);
2216         vfree(ca->buckets);
2217
2218         free_heap(&ca->heap);
2219         free_fifo(&ca->free_inc);
2220
2221         for (i = 0; i < RESERVE_NR; i++)
2222                 free_fifo(&ca->free[i]);
2223
2224         if (ca->sb_disk)
2225                 put_page(virt_to_page(ca->sb_disk));
2226
2227         if (!IS_ERR_OR_NULL(ca->bdev))
2228                 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2229
2230         kfree(ca);
2231         module_put(THIS_MODULE);
2232 }
2233
2234 static int cache_alloc(struct cache *ca)
2235 {
2236         size_t free;
2237         size_t btree_buckets;
2238         struct bucket *b;
2239         int ret = -ENOMEM;
2240         const char *err = NULL;
2241
2242         __module_get(THIS_MODULE);
2243         kobject_init(&ca->kobj, &bch_cache_ktype);
2244
2245         bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2246
2247         /*
2248          * when ca->sb.njournal_buckets is not zero, journal exists,
2249          * and in bch_journal_replay(), tree node may split,
2250          * so bucket of RESERVE_BTREE type is needed,
2251          * the worst situation is all journal buckets are valid journal,
2252          * and all the keys need to replay,
2253          * so the number of  RESERVE_BTREE type buckets should be as much
2254          * as journal buckets
2255          */
2256         btree_buckets = ca->sb.njournal_buckets ?: 8;
2257         free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2258         if (!free) {
2259                 ret = -EPERM;
2260                 err = "ca->sb.nbuckets is too small";
2261                 goto err_free;
2262         }
2263
2264         if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2265                                                 GFP_KERNEL)) {
2266                 err = "ca->free[RESERVE_BTREE] alloc failed";
2267                 goto err_btree_alloc;
2268         }
2269
2270         if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2271                                                         GFP_KERNEL)) {
2272                 err = "ca->free[RESERVE_PRIO] alloc failed";
2273                 goto err_prio_alloc;
2274         }
2275
2276         if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2277                 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2278                 goto err_movinggc_alloc;
2279         }
2280
2281         if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2282                 err = "ca->free[RESERVE_NONE] alloc failed";
2283                 goto err_none_alloc;
2284         }
2285
2286         if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2287                 err = "ca->free_inc alloc failed";
2288                 goto err_free_inc_alloc;
2289         }
2290
2291         if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2292                 err = "ca->heap alloc failed";
2293                 goto err_heap_alloc;
2294         }
2295
2296         ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2297                               ca->sb.nbuckets));
2298         if (!ca->buckets) {
2299                 err = "ca->buckets alloc failed";
2300                 goto err_buckets_alloc;
2301         }
2302
2303         ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2304                                    prio_buckets(ca), 2),
2305                                    GFP_KERNEL);
2306         if (!ca->prio_buckets) {
2307                 err = "ca->prio_buckets alloc failed";
2308                 goto err_prio_buckets_alloc;
2309         }
2310
2311         ca->disk_buckets = alloc_meta_bucket_pages(GFP_KERNEL, &ca->sb);
2312         if (!ca->disk_buckets) {
2313                 err = "ca->disk_buckets alloc failed";
2314                 goto err_disk_buckets_alloc;
2315         }
2316
2317         ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2318
2319         for_each_bucket(b, ca)
2320                 atomic_set(&b->pin, 0);
2321         return 0;
2322
2323 err_disk_buckets_alloc:
2324         kfree(ca->prio_buckets);
2325 err_prio_buckets_alloc:
2326         vfree(ca->buckets);
2327 err_buckets_alloc:
2328         free_heap(&ca->heap);
2329 err_heap_alloc:
2330         free_fifo(&ca->free_inc);
2331 err_free_inc_alloc:
2332         free_fifo(&ca->free[RESERVE_NONE]);
2333 err_none_alloc:
2334         free_fifo(&ca->free[RESERVE_MOVINGGC]);
2335 err_movinggc_alloc:
2336         free_fifo(&ca->free[RESERVE_PRIO]);
2337 err_prio_alloc:
2338         free_fifo(&ca->free[RESERVE_BTREE]);
2339 err_btree_alloc:
2340 err_free:
2341         module_put(THIS_MODULE);
2342         if (err)
2343                 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2344         return ret;
2345 }
2346
2347 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2348                                 struct block_device *bdev, struct cache *ca)
2349 {
2350         const char *err = NULL; /* must be set for any error case */
2351         int ret = 0;
2352
2353         bdevname(bdev, ca->cache_dev_name);
2354         memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2355         ca->bdev = bdev;
2356         ca->bdev->bd_holder = ca;
2357         ca->sb_disk = sb_disk;
2358
2359         if (blk_queue_discard(bdev_get_queue(bdev)))
2360                 ca->discard = CACHE_DISCARD(&ca->sb);
2361
2362         ret = cache_alloc(ca);
2363         if (ret != 0) {
2364                 /*
2365                  * If we failed here, it means ca->kobj is not initialized yet,
2366                  * kobject_put() won't be called and there is no chance to
2367                  * call blkdev_put() to bdev in bch_cache_release(). So we
2368                  * explicitly call blkdev_put() here.
2369                  */
2370                 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2371                 if (ret == -ENOMEM)
2372                         err = "cache_alloc(): -ENOMEM";
2373                 else if (ret == -EPERM)
2374                         err = "cache_alloc(): cache device is too small";
2375                 else
2376                         err = "cache_alloc(): unknown error";
2377                 goto err;
2378         }
2379
2380         if (kobject_add(&ca->kobj, bdev_kobj(bdev), "bcache")) {
2381                 err = "error calling kobject_add";
2382                 ret = -ENOMEM;
2383                 goto out;
2384         }
2385
2386         mutex_lock(&bch_register_lock);
2387         err = register_cache_set(ca);
2388         mutex_unlock(&bch_register_lock);
2389
2390         if (err) {
2391                 ret = -ENODEV;
2392                 goto out;
2393         }
2394
2395         pr_info("registered cache device %s\n", ca->cache_dev_name);
2396
2397 out:
2398         kobject_put(&ca->kobj);
2399
2400 err:
2401         if (err)
2402                 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2403
2404         return ret;
2405 }
2406
2407 /* Global interfaces/init */
2408
2409 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2410                                const char *buffer, size_t size);
2411 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2412                                          struct kobj_attribute *attr,
2413                                          const char *buffer, size_t size);
2414
2415 kobj_attribute_write(register,          register_bcache);
2416 kobj_attribute_write(register_quiet,    register_bcache);
2417 kobj_attribute_write(pendings_cleanup,  bch_pending_bdevs_cleanup);
2418
2419 static bool bch_is_open_backing(dev_t dev)
2420 {
2421         struct cache_set *c, *tc;
2422         struct cached_dev *dc, *t;
2423
2424         list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2425                 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2426                         if (dc->bdev->bd_dev == dev)
2427                                 return true;
2428         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2429                 if (dc->bdev->bd_dev == dev)
2430                         return true;
2431         return false;
2432 }
2433
2434 static bool bch_is_open_cache(dev_t dev)
2435 {
2436         struct cache_set *c, *tc;
2437
2438         list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2439                 struct cache *ca = c->cache;
2440
2441                 if (ca->bdev->bd_dev == dev)
2442                         return true;
2443         }
2444
2445         return false;
2446 }
2447
2448 static bool bch_is_open(dev_t dev)
2449 {
2450         return bch_is_open_cache(dev) || bch_is_open_backing(dev);
2451 }
2452
2453 struct async_reg_args {
2454         struct delayed_work reg_work;
2455         char *path;
2456         struct cache_sb *sb;
2457         struct cache_sb_disk *sb_disk;
2458         struct block_device *bdev;
2459 };
2460
2461 static void register_bdev_worker(struct work_struct *work)
2462 {
2463         int fail = false;
2464         struct async_reg_args *args =
2465                 container_of(work, struct async_reg_args, reg_work.work);
2466         struct cached_dev *dc;
2467
2468         dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2469         if (!dc) {
2470                 fail = true;
2471                 put_page(virt_to_page(args->sb_disk));
2472                 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2473                 goto out;
2474         }
2475
2476         mutex_lock(&bch_register_lock);
2477         if (register_bdev(args->sb, args->sb_disk, args->bdev, dc) < 0)
2478                 fail = true;
2479         mutex_unlock(&bch_register_lock);
2480
2481 out:
2482         if (fail)
2483                 pr_info("error %s: fail to register backing device\n",
2484                         args->path);
2485         kfree(args->sb);
2486         kfree(args->path);
2487         kfree(args);
2488         module_put(THIS_MODULE);
2489 }
2490
2491 static void register_cache_worker(struct work_struct *work)
2492 {
2493         int fail = false;
2494         struct async_reg_args *args =
2495                 container_of(work, struct async_reg_args, reg_work.work);
2496         struct cache *ca;
2497
2498         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2499         if (!ca) {
2500                 fail = true;
2501                 put_page(virt_to_page(args->sb_disk));
2502                 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2503                 goto out;
2504         }
2505
2506         /* blkdev_put() will be called in bch_cache_release() */
2507         if (register_cache(args->sb, args->sb_disk, args->bdev, ca) != 0)
2508                 fail = true;
2509
2510 out:
2511         if (fail)
2512                 pr_info("error %s: fail to register cache device\n",
2513                         args->path);
2514         kfree(args->sb);
2515         kfree(args->path);
2516         kfree(args);
2517         module_put(THIS_MODULE);
2518 }
2519
2520 static void register_device_async(struct async_reg_args *args)
2521 {
2522         if (SB_IS_BDEV(args->sb))
2523                 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2524         else
2525                 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2526
2527         /* 10 jiffies is enough for a delay */
2528         queue_delayed_work(system_wq, &args->reg_work, 10);
2529 }
2530
2531 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2532                                const char *buffer, size_t size)
2533 {
2534         const char *err;
2535         char *path = NULL;
2536         struct cache_sb *sb;
2537         struct cache_sb_disk *sb_disk;
2538         struct block_device *bdev;
2539         ssize_t ret;
2540         bool async_registration = false;
2541
2542 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2543         async_registration = true;
2544 #endif
2545
2546         ret = -EBUSY;
2547         err = "failed to reference bcache module";
2548         if (!try_module_get(THIS_MODULE))
2549                 goto out;
2550
2551         /* For latest state of bcache_is_reboot */
2552         smp_mb();
2553         err = "bcache is in reboot";
2554         if (bcache_is_reboot)
2555                 goto out_module_put;
2556
2557         ret = -ENOMEM;
2558         err = "cannot allocate memory";
2559         path = kstrndup(buffer, size, GFP_KERNEL);
2560         if (!path)
2561                 goto out_module_put;
2562
2563         sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2564         if (!sb)
2565                 goto out_free_path;
2566
2567         ret = -EINVAL;
2568         err = "failed to open device";
2569         bdev = blkdev_get_by_path(strim(path),
2570                                   FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2571                                   sb);
2572         if (IS_ERR(bdev)) {
2573                 if (bdev == ERR_PTR(-EBUSY)) {
2574                         dev_t dev;
2575
2576                         mutex_lock(&bch_register_lock);
2577                         if (lookup_bdev(strim(path), &dev) == 0 &&
2578                             bch_is_open(dev))
2579                                 err = "device already registered";
2580                         else
2581                                 err = "device busy";
2582                         mutex_unlock(&bch_register_lock);
2583                         if (attr == &ksysfs_register_quiet)
2584                                 goto done;
2585                 }
2586                 goto out_free_sb;
2587         }
2588
2589         err = "failed to set blocksize";
2590         if (set_blocksize(bdev, 4096))
2591                 goto out_blkdev_put;
2592
2593         err = read_super(sb, bdev, &sb_disk);
2594         if (err)
2595                 goto out_blkdev_put;
2596
2597         err = "failed to register device";
2598
2599         if (async_registration) {
2600                 /* register in asynchronous way */
2601                 struct async_reg_args *args =
2602                         kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2603
2604                 if (!args) {
2605                         ret = -ENOMEM;
2606                         err = "cannot allocate memory";
2607                         goto out_put_sb_page;
2608                 }
2609
2610                 args->path      = path;
2611                 args->sb        = sb;
2612                 args->sb_disk   = sb_disk;
2613                 args->bdev      = bdev;
2614                 register_device_async(args);
2615                 /* No wait and returns to user space */
2616                 goto async_done;
2617         }
2618
2619         if (SB_IS_BDEV(sb)) {
2620                 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2621
2622                 if (!dc)
2623                         goto out_put_sb_page;
2624
2625                 mutex_lock(&bch_register_lock);
2626                 ret = register_bdev(sb, sb_disk, bdev, dc);
2627                 mutex_unlock(&bch_register_lock);
2628                 /* blkdev_put() will be called in cached_dev_free() */
2629                 if (ret < 0)
2630                         goto out_free_sb;
2631         } else {
2632                 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2633
2634                 if (!ca)
2635                         goto out_put_sb_page;
2636
2637                 /* blkdev_put() will be called in bch_cache_release() */
2638                 if (register_cache(sb, sb_disk, bdev, ca) != 0)
2639                         goto out_free_sb;
2640         }
2641
2642 done:
2643         kfree(sb);
2644         kfree(path);
2645         module_put(THIS_MODULE);
2646 async_done:
2647         return size;
2648
2649 out_put_sb_page:
2650         put_page(virt_to_page(sb_disk));
2651 out_blkdev_put:
2652         blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2653 out_free_sb:
2654         kfree(sb);
2655 out_free_path:
2656         kfree(path);
2657         path = NULL;
2658 out_module_put:
2659         module_put(THIS_MODULE);
2660 out:
2661         pr_info("error %s: %s\n", path?path:"", err);
2662         return ret;
2663 }
2664
2665
2666 struct pdev {
2667         struct list_head list;
2668         struct cached_dev *dc;
2669 };
2670
2671 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2672                                          struct kobj_attribute *attr,
2673                                          const char *buffer,
2674                                          size_t size)
2675 {
2676         LIST_HEAD(pending_devs);
2677         ssize_t ret = size;
2678         struct cached_dev *dc, *tdc;
2679         struct pdev *pdev, *tpdev;
2680         struct cache_set *c, *tc;
2681
2682         mutex_lock(&bch_register_lock);
2683         list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2684                 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2685                 if (!pdev)
2686                         break;
2687                 pdev->dc = dc;
2688                 list_add(&pdev->list, &pending_devs);
2689         }
2690
2691         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2692                 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2693                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2694                         char *set_uuid = c->set_uuid;
2695
2696                         if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2697                                 list_del(&pdev->list);
2698                                 kfree(pdev);
2699                                 break;
2700                         }
2701                 }
2702         }
2703         mutex_unlock(&bch_register_lock);
2704
2705         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2706                 pr_info("delete pdev %p\n", pdev);
2707                 list_del(&pdev->list);
2708                 bcache_device_stop(&pdev->dc->disk);
2709                 kfree(pdev);
2710         }
2711
2712         return ret;
2713 }
2714
2715 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2716 {
2717         if (bcache_is_reboot)
2718                 return NOTIFY_DONE;
2719
2720         if (code == SYS_DOWN ||
2721             code == SYS_HALT ||
2722             code == SYS_POWER_OFF) {
2723                 DEFINE_WAIT(wait);
2724                 unsigned long start = jiffies;
2725                 bool stopped = false;
2726
2727                 struct cache_set *c, *tc;
2728                 struct cached_dev *dc, *tdc;
2729
2730                 mutex_lock(&bch_register_lock);
2731
2732                 if (bcache_is_reboot)
2733                         goto out;
2734
2735                 /* New registration is rejected since now */
2736                 bcache_is_reboot = true;
2737                 /*
2738                  * Make registering caller (if there is) on other CPU
2739                  * core know bcache_is_reboot set to true earlier
2740                  */
2741                 smp_mb();
2742
2743                 if (list_empty(&bch_cache_sets) &&
2744                     list_empty(&uncached_devices))
2745                         goto out;
2746
2747                 mutex_unlock(&bch_register_lock);
2748
2749                 pr_info("Stopping all devices:\n");
2750
2751                 /*
2752                  * The reason bch_register_lock is not held to call
2753                  * bch_cache_set_stop() and bcache_device_stop() is to
2754                  * avoid potential deadlock during reboot, because cache
2755                  * set or bcache device stopping process will acqurie
2756                  * bch_register_lock too.
2757                  *
2758                  * We are safe here because bcache_is_reboot sets to
2759                  * true already, register_bcache() will reject new
2760                  * registration now. bcache_is_reboot also makes sure
2761                  * bcache_reboot() won't be re-entered on by other thread,
2762                  * so there is no race in following list iteration by
2763                  * list_for_each_entry_safe().
2764                  */
2765                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2766                         bch_cache_set_stop(c);
2767
2768                 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2769                         bcache_device_stop(&dc->disk);
2770
2771
2772                 /*
2773                  * Give an early chance for other kthreads and
2774                  * kworkers to stop themselves
2775                  */
2776                 schedule();
2777
2778                 /* What's a condition variable? */
2779                 while (1) {
2780                         long timeout = start + 10 * HZ - jiffies;
2781
2782                         mutex_lock(&bch_register_lock);
2783                         stopped = list_empty(&bch_cache_sets) &&
2784                                 list_empty(&uncached_devices);
2785
2786                         if (timeout < 0 || stopped)
2787                                 break;
2788
2789                         prepare_to_wait(&unregister_wait, &wait,
2790                                         TASK_UNINTERRUPTIBLE);
2791
2792                         mutex_unlock(&bch_register_lock);
2793                         schedule_timeout(timeout);
2794                 }
2795
2796                 finish_wait(&unregister_wait, &wait);
2797
2798                 if (stopped)
2799                         pr_info("All devices stopped\n");
2800                 else
2801                         pr_notice("Timeout waiting for devices to be closed\n");
2802 out:
2803                 mutex_unlock(&bch_register_lock);
2804         }
2805
2806         return NOTIFY_DONE;
2807 }
2808
2809 static struct notifier_block reboot = {
2810         .notifier_call  = bcache_reboot,
2811         .priority       = INT_MAX, /* before any real devices */
2812 };
2813
2814 static void bcache_exit(void)
2815 {
2816         bch_debug_exit();
2817         bch_request_exit();
2818         if (bcache_kobj)
2819                 kobject_put(bcache_kobj);
2820         if (bcache_wq)
2821                 destroy_workqueue(bcache_wq);
2822         if (bch_journal_wq)
2823                 destroy_workqueue(bch_journal_wq);
2824         if (bch_flush_wq)
2825                 destroy_workqueue(bch_flush_wq);
2826         bch_btree_exit();
2827
2828         if (bcache_major)
2829                 unregister_blkdev(bcache_major, "bcache");
2830         unregister_reboot_notifier(&reboot);
2831         mutex_destroy(&bch_register_lock);
2832 }
2833
2834 /* Check and fixup module parameters */
2835 static void check_module_parameters(void)
2836 {
2837         if (bch_cutoff_writeback_sync == 0)
2838                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2839         else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2840                 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2841                         bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2842                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2843         }
2844
2845         if (bch_cutoff_writeback == 0)
2846                 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2847         else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2848                 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2849                         bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2850                 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2851         }
2852
2853         if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2854                 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2855                         bch_cutoff_writeback, bch_cutoff_writeback_sync);
2856                 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2857         }
2858 }
2859
2860 static int __init bcache_init(void)
2861 {
2862         static const struct attribute *files[] = {
2863                 &ksysfs_register.attr,
2864                 &ksysfs_register_quiet.attr,
2865                 &ksysfs_pendings_cleanup.attr,
2866                 NULL
2867         };
2868
2869         check_module_parameters();
2870
2871         mutex_init(&bch_register_lock);
2872         init_waitqueue_head(&unregister_wait);
2873         register_reboot_notifier(&reboot);
2874
2875         bcache_major = register_blkdev(0, "bcache");
2876         if (bcache_major < 0) {
2877                 unregister_reboot_notifier(&reboot);
2878                 mutex_destroy(&bch_register_lock);
2879                 return bcache_major;
2880         }
2881
2882         if (bch_btree_init())
2883                 goto err;
2884
2885         bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2886         if (!bcache_wq)
2887                 goto err;
2888
2889         /*
2890          * Let's not make this `WQ_MEM_RECLAIM` for the following reasons:
2891          *
2892          * 1. It used `system_wq` before which also does no memory reclaim.
2893          * 2. With `WQ_MEM_RECLAIM` desktop stalls, increased boot times, and
2894          *    reduced throughput can be observed.
2895          *
2896          * We still want to user our own queue to not congest the `system_wq`.
2897          */
2898         bch_flush_wq = alloc_workqueue("bch_flush", 0, 0);
2899         if (!bch_flush_wq)
2900                 goto err;
2901
2902         bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2903         if (!bch_journal_wq)
2904                 goto err;
2905
2906         bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2907         if (!bcache_kobj)
2908                 goto err;
2909
2910         if (bch_request_init() ||
2911             sysfs_create_files(bcache_kobj, files))
2912                 goto err;
2913
2914         bch_debug_init();
2915         closure_debug_init();
2916
2917         bcache_is_reboot = false;
2918
2919         return 0;
2920 err:
2921         bcache_exit();
2922         return -ENOMEM;
2923 }
2924
2925 /*
2926  * Module hooks
2927  */
2928 module_exit(bcache_exit);
2929 module_init(bcache_init);
2930
2931 module_param(bch_cutoff_writeback, uint, 0);
2932 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2933
2934 module_param(bch_cutoff_writeback_sync, uint, 0);
2935 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2936
2937 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2938 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2939 MODULE_LICENSE("GPL");